Letter doi:10.

1038/nature18326

Anthropogenic disturbance in tropical forests can

double biodiversity loss from deforestation
Jos Barlow1,2,3, Gareth D. Lennox1, Joice Ferreira4, Erika Berenguer1, Alexander C. Lees2,5, Ralph Mac Nally6,
James R. Thomson6,7, Silvio Frosini de Barros Ferraz8, Julio Louzada1,3, Victor Hugo Fonseca Oliveira1,3, Luke Parry1,9,
Ricardo Ribeiro de  Castro Solar10, Ima C. G. Vieira2, Luiz E. O. C. Aragão11,12, Rodrigo Anzolin Begotti8, Rodrigo F. Braga3,
Thiago Moreira Cardoso4, Raimundo Cosme de Oliveira Jr4, Carlos M. Souza Jr13, Nárgila G. Moura2,5, Sâmia Serra Nunes13,
João Victor Siqueira13, Renata Pardini14, Juliana M. Silveira1,3, Fernando Z. Vaz-de-Mello15, Ruan Carlo Stulpen Veiga16,
Adriano Venturieri4 & Toby A. Gardner17,18

Concerted political attention has focused on reducing
deforestation 1–3, and this remains the cornerstone of most
biodiversity conservation strategies4–6. However, maintaining
forest cover may not reduce anthropogenic forest disturbances,
which are rarely considered in conservation programmes6. These
disturbances occur both within forests, including selective logging
and wildfires7,8, and at the landscape level, through edge, area and
isolation effects9. Until now, the combined effect of anthropogenic
disturbance on the conservation value of remnant primary forests
has remained unknown, making it impossible to assess the relative
importance of forest disturbance and forest loss. Here we address
these knowledge gaps using a large data set of plants, birds and
dung beetles (1,538, 460 and 156 species, respectively) sampled in
36 catchments in the Brazilian state of Pará. Catchments retaining
more than 69–80% forest cover lost more conservation value from
disturbance than from forest loss. For example, a 20% loss of primary
forest, the maximum level of deforestation allowed on Amazonian
properties under Brazil’s Forest Code5, resulted in a 39–54% loss
of conservation value: 96–171% more than expected without
considering disturbance effects. We extrapolated the disturbance-
mediated loss of conservation value throughout Pará, which covers
25% of the Brazilian Amazon. Although disturbed forests retained
considerable conservation value compared with deforested areas,
the toll of disturbance outside Pará’s strictly protected areas is
equivalent to the loss of 92,000–139,000 km2 of primary forest. Even
this lowest estimate is greater than the area deforested across the
entire Brazilian Amazon between 2006 and 2015 (ref. 10). Species
distribution models showed that both landscape and within-forest
disturbances contributed to biodiversity loss, with the greatest
negative effects on species of high conservation and functional value.
These results demonstrate an urgent need for policy interventions
that go beyond the maintenance of forest cover to safeguard the
hyper-diversity of tropical forest ecosystems.
Protecting tropical forests is a fundamental pillar of many national
and international strategies for conserving biodiversity4–6. Although
improved regulatory and incentive measures have reduced deforesta-
tion rates in some tropical nations1,11,12, the conservation value of the
world’s remaining primary forests may be undermined by the addi-
tional impacts of disturbance, which falls into two broad categories (see
Methods). First, landscape disturbance results from deforestation itself,
with area, isolation and edge effects degrading the condition of the
remaining forests9. Second, within-forest disturbance, such as wildfires
and selective logging, induces marked changes in forest structure and
species composition8,13.
Although the biodiversity consequences of both forms of distur-
bance are well studied, previous research has overwhelmingly focused
on identifying the isolated effects of specific types of disturbance14,15.
Such studies provide an incomplete understanding of the total
disturbance-mediated loss of conservation value arising from multiple
interacting drivers16 and are unable to quantify the extent to which
reducing forest loss will succeed in protecting tropical forest biodiver-
sity. Addressing these knowledge gaps is vital for informing forest man-
agement strategies in tropical nations, not least because within-forest
disturbance can increase even as deforestation rates fall7,12,17 and thus
requires different policy interventions (Extended Data Table 1).
We estimated the combined effects of landscape and within-forest
disturbance on biodiversity in primary forests and compared these
impacts to the biodiversity loss expected in deforested areas, offering,
to our knowledge, the first such analysis for anywhere in the world.
Our study focused on two large (>​10,000  km2) frontier regions of the
Brazilian Amazon: Paragominas and Santarém, located in the state of
Pará (see Methods). Large- and small-stemmed plants, birds and dung
beetles were sampled in 371 plots in 36 study catchments distributed
along a landscape deforestation gradient (0–94%) (Extended Data
Fig. 1). A total of 31 catchments contained remnant primary forests.
Within these catchments, we sampled 175 primary forest plots. Of
these, 145 had visible evidence of within-forest disturbance (logging
and/or fire). The remaining 30 had no evidence of within-forest dis-
turbance and, being located in the largest remaining forest blocks, had
minimal landscape disturbance18,19 (see Methods). Irrespective of their
disturbance history, these primary forest plots held considerably more
forest species than all other major land-uses (Extended Data Fig. 2).
We used the sum of forest species presences in primary forest plots to
estimate the conservation value of a catchment (see Methods). As plots

1
Lancaster Environment Centre, Lancaster University, Lancaster LA1 4YQ, UK. 2MCTI/Museu Paraense Emílio Goeldi, CP 399, Belém, Pará, CEP 66040-170, Brazil. 3Universidade Federal de Lavras,
Setor de Ecologia e Conservação. Lavras, Minas Gerais, CEP 37200-000, Brazil. 4EMBRAPA Amazônia Oriental. Belém, Pará, CEP 66095-100, Brazil. 5Cornell Lab of Ornithology, Cornell University,
Ithaca, New York 14850, USA. 6Institute for Applied Ecology, University of Canberra, Bruce, Australian Capital Territory 2617, Australia. 7Arthur Rylah Institute for Environmental Research,
Department of Environment, Land, Water and Planning, 123 Brown Street, Heidelberg, Victoria 3084, Australia. 8Universidade de São Paulo, Escola Superior de Agricultura “Luiz de Queiroz”,
Esalq/USP, Avenida Pádua Dias, 11, São Dimas, Piracicaba, SP, CEP 13418-900, Brazil. 9Universidade Federal do Pará (UFPA), Núcleo de Altos Estudos Amazonicos (NAEA), Av. Perimetral, Numero 1,
Guamá, Belém-Pará, CEP 66075-750, Brazil. 10Universidade Federal de Viçosa, Departamento de Biologia Geral. Av. PH Rolfs s/n. Viçosa, Minas Gerais, CEP 36570-900, Brazil. 11Tropical
Ecosystems and Environmental Sciences Group (TREES), Remote Sensing Division, National Institute for Space Research (INPE), Avenida dos Astronautas, 1.758, Jd. Granja, São José dos Campos,
CEP 12227-010, SP, Brazil. 12College of Life and Environmental Sciences, University of Exeter, Exeter EX4 4RJ, UK. 13IMAZON, Rua Dom Romualdo de Seixas 1698, Edifício Zion, 11 andar, CEP
66055-200 Belém, PA, Brazil. 14Instituto de Biociencias, Universidade de São Paulo, Rua do Matão, Travessa 14, 101, CEP 05508-090 São Paulo, Brazil. 15Universidade Federal de Mato Grosso,
Instituto de Biociencias, Departamento de Biologia e Zoologia. Av. Fernando Correa da Costa, 2367, Boa Esperança, CEP 78060-900, Cuiaba, MT, Brazil. 16Instituto Socio Ambiental Serra do
Mar (ISASM), Estrada Ribeirão das Voltas s/n, Lumiar, CEP 28616-010, Nova Friburgo, Brazil. 17Stockholm Environment Institute, Linnégatan 87D, Box 24218, Stockholm 104 51, Sweden.
18
International Institute for Sustainability, Estrada Dona Castorina, 124, Horto, Rio de Janeiro, 22460-320, Brazil.

0 0 M o n t h 2 0 1 6 | VO L 0 0 0 | NAT U R E | 1
© 2016 Macmillan Publishers Limited. All rights reserved
 RESEARCH Letter

a
a 0 200 400 km
1.0

0.8 GU
MA
Conservation value

BE
0.6
RO
Loss
of CV
0.4 XI
TA
0.2

0.2 0.1

0.0
0.0

b 0.8 Forest loss 13–23

b Disturbance
0.4
20–33
0.6

Loss of conservation value

Absolute loss (CVD)

0.3
38–57

0.4
92–137
0.2
106–151
135–178

0.1 0.2

0.0
0.0

PA TA GU XI BE RO
c Figure 2 | Conservation value deficit over large spatial scales.
0.6
a, Proportionate loss of conservation value (CV) from disturbance in
Pará (median estimate; see Methods). Areas of endemism (AoE) are:
Belém (BE), Guiana (GU), Rondônia (RO), Tapajós (TA) and Xingu (XI).
These do not include the island of Marajó (MA). Grey shading denotes
Proportionate loss

0.4
strictly protected areas. b, Proportionate loss of CV in Pará (PA) and its
AoEs from forest loss and disturbance (median estimate). Error bars
show the range over all approaches to estimating conservation value
0.2
(see Methods). Numbers show disturbance relative to forest loss
(percentage range over approaches).

0.0
0.00 0.25 0.50 0.75
Primary forest cover
Figure 1 | The conservation status of primary forests. a, Conservation
value in Paragominas (circles) and Santarém (triangles). b, Total loss of
conservation value due to disturbance. c, Total loss of conservation value
due to disturbance expressed as a proportion of the expected conservation
value without disturbance. Dashed lines show expectations without
disturbance. Grey lines show all regressions, with the black solid line
showing the median response (see Methods). Values were standardized
across study regions. There was no significant difference in conservation
values between regions in the median response (F1;26 =​  1.45, P =​  0.24,
analysis of covariance (ANCOVA)).
were allocated in proportion to catchment forest cover, this measure is
equivalent to the mean species richness (per unit area) in primary for-
ests multiplied by the proportion of primary forest cover. In the absence
of landscape or within-forest disturbances, the expectation of conser-
vation value should respond linearly to forest cover, with slope equal to
mean species density (see Methods). The difference between this linear
expectation and the observed conservation value of the remaining pri-
mary forest provides an estimate of the total biodiversity impact of all
landscape and within-forest disturbance. We refer to this difference as
the conservation value deficit (CVD). We take a variety of approaches
to calculating the CVD, reflecting different ways of classifying forest
species, weighting their conservation value and calculating species den-
1.00 sity in undisturbed forest (see Methods). Here, we report median results
from our sensitivity analysis along with the lower and upper bound
range. Full results are shown in Fig. 1 and Extended Data Figs 3 and 4.
The conservation value of the remaining primary forests was lower
than expected along the entire deforestation gradient. The CVD was
unimodal with forest cover, reaching its maximum in catchments with
83% of their primary forests. These catchments retained just 58% of
their conservation value (range: 48 to 65%) (Fig. 1a). The CVD was
relatively small at low levels of forest cover (Fig. 1b). However, distur-
bance caused the greatest proportionate loss of conservation value in
these catchments, accounting for an approximately 20–50% shortfall in
the level of biodiversity that would be predicted for undisturbed forests
(Fig. 1c). The robustness of our estimates of the CVD was supported
by the similarity of responses across study regions (Fig. 1) and sampled
taxa (Extended Data Fig. 3).
The relationship we derived between forest cover and conservation
value allowed us, for the first time, to estimate the additional total effect
of forest disturbance over large spatial scales. We therefore mapped
the disturbance-induced loss of conservation value (CVD) across
Pará, which covers 1.26 × 106 km2. We divided the state into grid cells
approximately equal in area to our study catchments (~​50  km2). In
total, 73% of the ~​26,000 cells covering the state were located in private
lands or sustainable-use reserves. For these locations, which are most

2 | NAT U R E | VO L 0 0 0 | 0 0 m o n t h 2 0 1 6
© 2016 Macmillan Publishers Limited. All rights reserved
 Letter RESEARCH

Landscape disturbance Within-forest disturbance Figure 3 | Response of forest birds to

a b c d disturbance. a–d, The odds of detecting
Paragominas Santarém Paragominas Santarém
4.0
2.5 species groups along gradients of
2.0
2.0 landscape (a, b) and within-forest (c, d)
2.0
species detection

disturbance in Paragominas (a, c) and

Relative odds of

3.0
1.5 1.5 1.5 Santarém (b, d) (see Methods). Species
2.0 groups, shown by different coloured lines,
1.0 1.0 1.0 are composed of species with similar
1.0 disturbance responses (see Methods). Line
0.5
0.5 0.5 thickness represents the relative size of
25 50 75 100 20 40 60 80 100 40 60 80 100 40 60 80 100 the groups. e–h, Disturbance sensitivity
Forest cover Forest cover Undegraded forest Undegraded forest of the species groups related to their
mean range size (107 km2). Error bars
e f g h shows s.e.m. Group colours correspond
to groupings in a–d. Black lines show
1.5 1.5 1.5 1.5 significant relationships (P <​  0.05, F-test)
(see Methods).
Meanrange size

●
1.2 1.2 1.2 1.2
of birds

0.9 0.9 0.9 0.9

0.6 0.6 0.6 0.6

0.3 0.3 0.3 0.3

−1.0 −0.5 0.0 0.5 −0.8 −0.4 0.0 0.4 −1.0 −0.5 0.0 0.5 −1.0 −0.5 0.0 0.5 1.0

Low High Low High Low High Low High

Disturbance sensitivity of response groups

comparable to our study catchments, the total CVD was equivalent to conservation value from disturbance was 135–178% of the losses
~​123,000  km2 of forest loss (range: 92,000 to 139,000 km2). To put this estimated in deforested areas. The relative impact of disturbance was
figure in context, it is 51% (range: 38 to 57%) of the total area deforested lowest in the Belém AoE, which has lost 62% of its native forest cover
across Pará to date (Extended Data Table 2). and is the most deforested AoE in Amazonia. Nonetheless, overall
Our state-wide analysis revealed considerable spatial variation in disturbance effects reduced Belém’s estimated conservation value
the CVD, reflecting differences in deforestation histories (Fig. 2a). from 38% when based on forest cover alone to just 26% (range: 24
We illustrate this variation by estimating the additional loss of con- to 30%).
servation value due to disturbance across Pará’s five major biogeo­ The widespread and substantial depletion of conservation value in
graphic zones (areas of endemism20, AoE). Median disturbance remaining primary forests highlights the pressing need for policies that
impacts outweighed biodiversity losses in deforested areas alone target the most prominent drivers of disturbance-induced biodiver-
in three of the five AoEs (Fig. 2b). The high relative impact of dis- sity loss. Although measures to combat deforestation may help limit
turbance is shown in the Guiana AoE, where the predicted loss of landscape disturbance, they rarely consider the spatial configuration of

Landscape disturbance Within-forest disturbance

a Paragominas
b Santarém
c Paragominas d Santarém
2.4
2.0 3.0
species detection
Relative odds of

2.0
2.0

1.5 1.5 1.6 2.0

1.2
1.0 1.0 1.0
0.8

25 50 75 100 20 40 60 80 100 40 60 80 100 40 60 80 100

Forest cover Forest cover Undegraded forest Undegraded forest
e f g h

0.8 0.8 0.8 0.8

Mean wood density
of large stems

0.6 0.6 0.6 0.6

0.4 0.4 0.4 0.4

−0.5 0.0 0.5 1.0 −1.0 −0.5 0.0 0.5 −1.0 −0.5 0.0 0.5 1.0 −1.0 −0.5 0.0 0.5

Low High Low High Low High Low High

Disturbance sensitivity of response groups

Figure 4 | Response of large-stemmed plants to disturbance. a–d, The Line thickness represents the relative size of the groups. e–h, Disturbance
odds of detecting species groups along gradients of landscape (a, b) and sensitivity of the species groups related to their mean wood density
within-forest (c, d) disturbance in Paragominas (a, c) and Santarém (b, d) (g cm−3). Error bars show s.e.m. Group colours correspond to groupings
(see Methods). Species groups, shown by different coloured lines, are in a–d. Black lines show significant relationships (P <​  0.05, F-test)
composed of species with similar disturbance responses (see Methods). (see Methods).

0 0 M o n t h 2 0 1 6 | VO L 0 0 0 | NAT U R E | 3
© 2016 Macmillan Publishers Limited. All rights reserved
 RESEARCH Letter
remaining forests or work to actively reduce within-forest disturbance 6,21
(Extended Data Table 1).
Here we provide insights into the need for additional policies to
reduce forest disturbance by examining the relative importance of
landscape and within-forest disturbance on species distributions
using Random Forests (see Methods). In ranking the importance of
remotely sensed disturbance measures, we found that both forms of
disturbance had significant additional effects on species’ distributions,
albeit with some region- and taxon-specific variation (Extended Data
Figs 5–7 and Methods). We then used the measures of landscape and
within-forest disturbance that were most frequently ranked high-
est to examine changes in taxon community structure, using Latent
Trajectory Analysis to group species by their responses to disturbance
(see Methods). Results showed a consistent and high level of com-
munity turnover from both forms of disturbance, with some species
groups responding negatively and others positively (Figs 3 and 4 and
Extended Data Fig. 8). These responses may explain the unimodal
shape of the disturbance effect (Fig. 1b) because they are consistent
with the loss of highly sensitive species at relatively low levels of forest
disturbance and the dominance of more resistant taxa in the most
disturbed forests. Finally, we linked species’ response groups with life-
history data available for birds and large-stemmed plants (see
Methods). Both types of disturbance contributed to marked declines
in species of high conservation and functional importance (birds
with smaller range sizes22,23 and plants with higher wood density24–26,
respectively) (Figs 3 and 4). These analyses almost certainly underesti-
mate the adverse effects of disturbance because rare species, which are
often most sensitive to human impacts in forest ecosystems27, cannot
be adequately modelled.
We provide compelling evidence that Amazonian conservation ini-
tiatives must address forest disturbance as well as deforestation. At its
most stringent, Brazil’s centrepiece environmental legislation, the Forest
Code, mandates Amazonian landowners to maintain 80% of their pri-
mary forest cover. Our results show that even where this level of com-
pliance is achieved, the primary forests of these landscapes may only
retain 46–61% of their potential conservation value and are likely to
have lost many species of high conservation and functional importance.
These findings reinforce the need to reduce the effects of landscape
fragmentation by zoning development activities, thereby ensuring
the protection of large blocks of remaining forest in all biogeographic
zones. Where deforestation has already occurred, further conservation
losses can be minimised by preventing within-forest disturbance, aiding
the recovery of already degraded forests, and investing in forest resto-
ration to improve connectivity and buffer remnant forests from edge
effects. Engendering change will require a mixture of incentive and
regulatory-based measures to improve the sustainability of both forestry
and farming practices. Crucially, because reducing forest disturbance
requires coordinated efforts by many actors, interventions need to
move beyond individual properties and address entire landscapes and
regions. Such actions are urgently needed in the Amazon where log-
ging operations are rapidly expanding across federal and state forests28,
wildfires are increasingly prevalent during more frequent and severe dry
seasons29, and the expansion of industrial agriculture, energy and mining
threaten even strictly protected areas and indigenous lands30.
Online Content Methods, along with any additional Extended Data display items and
Source Data, are available in the online version of the paper; references unique to
these sections appear only in the online paper.
received 3 March; accepted 17 May 2016.
Published online 29 June 2016.
1. Boucher, D., Elias, P., Faires, J. & Smith, S. Deforestation Success Stories:
Tropical Nations Where Forest Protection and Reforestation Policies Have
Worked. Union of Concerned Scientists June 2014 Report (2014).
2. Nepstad, D. et al. The end of deforestation in the Brazilian Amazon. Science
326, 1350–1351 (2009).
3. Soares-Filho, B. S. et al. Modelling conservation in the Amazon basin.
Nature 440, 520–523 (2006).
4 | NAT U R E | VO L 0 0 0 | 0 0 m o n t h 2 0 1 6
© 2016 Macmillan Publishers Limited. All rights reserved
4. Convention on Biological Diversity. Strategic Plan for Biodiversity 2011–2020,
Aichi Biodiversity Targets https://www.cbd.int/sp/default.shtml (2015).
5. Legislative Database of the Food and Agricultural Organization of the United
Nations (FAOLEX). Brazilian Environmental Law number 12.651 (25 March 2012).
6. Panfil, S. N. & Harvey, C. A. REDD+​and Biodiversity Conservation: A review of
the biodiversity goals, monitoring methods and impacts of 80 REDD+​ projects.
Conserv. Lett. 9, 143–150 (2015).
7. Aragão, L. E. O. C. & Shimabukuro, Y. E. The incidence of fire in Amazonian
forests with implications for REDD. Science 328, 1275–1278 (2010).
8. Burivalova, Z., Ş   ekercioğlu, C. H. & Koh, L. P. Thresholds of logging intensity to
maintain tropical forest biodiversity. Curr. Biol. 24, 1893–1898 (2014).
9. Ewers, R. M. & Didham, R. K. Confounding factors in the detection of species
responses to habitat fragmentation. Biol. Rev. Camb. Philos. Soc. 81, 117–142
(2006).
10. Instituto Nacional de Pesquisas Espaciais (INPE). Projeto Prodes: Amazon
deforestation database. Available at www.obt.inpe.br/prodes (2015).
11. Hansen, M. C. et al. High-resolution global maps of 21st-century forest cover
change. Science 342, 850–853 (2013).
12. Sloan, S. & Sayer, J. Forest Resources Assessment of 2015 shows positive
global trends but forest loss and degradation persist in poor tropical countries.
For. Ecol. Manage. 352, 134–145 (2015).
13. Barlow, J. & Peres, C. A. Avifaunal responses to single and recurrent wildfires in
Amazonian forests. Ecol. Appl. 14, 1358–1373 (2004).
14. Lewis, S. L., Edwards, D. P. & Galbraith, D. Increasing human dominance of
tropical forests. Science 349, 827–832 (2015).
15. Gibson, L. et al. Primary forests are irreplaceable for sustaining tropical
biodiversity. Nature 478, 378–381 (2011).
16. Malhi, Y., Gardner, T. A., Goldsmith, G. R., Silman, M. R. & Zelazowski, P. Tropical
Forests in the Anthropocene. Annu. Rev. Environ. Resour. 39, 125–159 (2014).
17. Morton, D. C., Le Page, Y., DeFries, R., Collatz, G. J. & Hurtt, G. C. Understorey
fire frequency and the fate of burned forests in southern Amazonia. Phil. Trans.
R. Soc. B 368, 1–8 (2013).
18. Gardner, T. A. et al. A social and ecological assessment of tropical land uses at
multiple scales: the Sustainable Amazon Network. Phil. Tran. R. Soc. B 368,
20120166 (2013).
19. Berenguer, E. et al. A large-scale field assessment of carbon stocks in
human-modified tropical forests. Glob. Chang. Biol. 20, 3713–3726 (2014).
20. da Silva, J. M. C., Rylands, A. B. & Da Fonseca, G. A. B. The fate of the
Amazonian areas of endemism. Conserv. Biol. 19, 689–694 (2005).
21. International Union of Forest Research Organizations (IUFRO). Understanding
Relationships between Biodiversity, Carbon, Forests and People: The Key to
Achieving REDD+ Objectives (eds Parrotta, J. A., Wildburger, C. & Mansourian, S.)
(2012).
22. Manne, L. L., Brooks, T. M. & Pimm, S. L. Relative risk of extinction of passerine
birds on continents and islands. Nature 399, 258–261 (1999).
23. Purvis, A., Gittleman, J. L., Cowlishaw, G. & Mace, G. M. Predicting extinction
risk in declining species. Proc. Biol. Sci. 267, 1947–1952 (2000).
24. Chave, J. et al. Towards a worldwide wood economics spectrum. Ecol. Lett. 12,
351–366 (2009).
25. Phillips, O. L. et al. Drought sensitivity of the Amazon rainforest. Science 323,
1344–1347 (2009).
26. Baker, T. R. et al. Variation in wood density determines spatial patterns in
Amazonian forest biomass. Glob. Chang. Biol. 10, 545–562 (2004).
27. Banks-Leite, C. et al. Assessing the utility of statistical adjustments for imperfect
detection in tropical conservation science. J. Appl. Ecol. 51, 849–859 (2014).
28. Gestão de Florestas Públicas – Relatório 2015. Brasília: MMA/SFB Available at
http://www.florestal.gov.br/publicacoes/instrumento-de-gestao (2015).
29. Chen, Y. et al. Forecasting fire season severity in South America using sea
surface temperature anomalies. Science 334, 787–791 (2011).
30. Ferreira, J. et al. Environment and Development. Brazil’s environmental
leadership at risk. Science 346, 706–707 (2014).
Acknowledgements This work was supported by grants from Brazil
(CNPq 574008/2008-0, 458022/2013-6, and 400640/2012-0; Embrapa
SEG:02.08.06.005.00; The Nature Conservancy – Brasil; CAPES scholarships)
the UK (Darwin Initiative 17-023; NE/F01614X/1; NE/G000816/1; NE/
F015356/2; NE/l018123/1; NE/K016431/1), Formas 2013-1571, and
Australian Research Council grant DP120100797. Institutional support was
provided by the Herbário IAN in Belém, LBA in Santarém and FAPEMAT. R.M.
and J.R.T. were supported by Australian Research Council grant DP120100797.
This is paper no. 49 in the Sustainable Amazon Network series.
Author Contributions T.A.G., J.F. and J.B. designed the research with additional
input from E.B., A.C.L., S.F.B.F., J.L., V.H.F.O., L.P., R.R.C.S., I.C.G.V., L.E.O.C.A. and
R.P. E.B., A.C.L., V.H.F.O., R.R.C.S, R.F.B., J.F., R.C.O., N.G.M. R.C.S.V., J.L., J.M.S
and F.Z.V. collected the field data or analysed biological or soil samples. G.D.L.
analysed the data, with input from J.B., J.R.T., R.M., A.C.L. and T.A.G. S.F.B.F.,
R.A.B., T.M.C., C.M.S., S.S.N., J.V.S., A.V. and T.A.G. processed the remote sensing
data. J.B., G.D.L., J.F., A.C.L., R.M., J.R.T. and T.A.G. wrote the manuscript, with
input from all authors.
Author Information Reprints and permissions information is available at
www.nature.com/reprints. The authors declare no competing financial
interests. Readers are welcome to comment on the online version of the
paper. Correspondence and requests for materials should be addressed to
J.B. (josbarlow@gmail.com).

Methods
No statistical methods were used to predetermine sample size. The experiments
were not randomized, and investigators were not blinded to allocation during
experiments and outcome assessment.
Study regions. Pará is the second largest state in Brazil and a focal point for
deforestation, accounting for 34% of all forest loss in the Brazilian Amazon between
1988 and 2015 (ref. 10). It holds exceptionally high biodiversity, with ~​10% of
the world’s bird species and five of the eight major AoEs in Amazonia20. Within
Pará, we focused on two geographically and biologically distinct regions: the
municipalities of Paragominas and Santarém-Belterra-Mojuí dos Campos (abbre-
viated to Santarém) (Extended Data Fig. 1). These regions lie in different AoEs
(Belém and Tapajós, respectively) and shared just 49% of our sampled taxa.
Although they differ in their human colonization history18, both retain >​50% of
their native forest cover.
Study design and biodiversity sampling. We divided each region into third- or
fourth-order drainage catchments. In each region, 18 study catchments
(32–61 km2) were then distributed along forest cover gradients. We distributed
study plots on terra firme in proportion to forest and non-forest cover at a density
of approximately 1 plot per 4 km2, resulting in 8–12 plots separated by ≥​1.5  km
in each catchment (Extended Data Fig. 1). Forest plots (n =​ 234) were distributed
without previous knowledge of anthropogenic disturbance18 and included pri-
mary forests (that is, under permanent forest cover; n =​ 175) and secondary forests
recovering after agricultural abandonment (n =​ 59). Non-forest plots (n =​  133)
were predominantly located in pastures (n =​ 76) and mechanised agricultural
lands (n =​  31).
In total, 31 of the 36 catchments contained primary forest plots. In Paragominas
and Santarém respectively, these included undisturbed (13 and 17), logged (44
and 26), burned (0 and 7) and logged and burned primary forests (44 and 24)19.
Disturbance categories were based on field assessments of fire scars, charcoal and
logging debris, and an analysis of canopy disturbance, deforestation and regrowth
in time series satellite images (1988 to 2010)18,19. Plots in the undisturbed forest had
no evidence of within-forest disturbance and, because they were located more than
2 km from edges in the largest forest blocks, had minimal landscape disturbance.
Observations of hunting-sensitive large game birds, such as razor-billed curassow
Pauxi tuberosa and trumpeters Psophia spp.31,32, indicated low hunting pressure33
in undisturbed plots.
Biodiversity surveys occurred during 2010 and 2011. The following descrip-
tions apply to sampling at the plot level. Large and small stems: live trees and
palms with ≥​10 cm diameter at breast height were identified in 10 ×​ 250 m plots.
Smaller individuals (2–10 cm diameter) were sampled in five 5 ×​ 20 m subplots
(Extended Data Fig. 1). Liana diameters were measured at 1.3 m from the main
root. Large- and small-stemmed plants were analysed separately because they may
differ in their disturbance responses. Individuals were identified to species level
by local parabotanists19. In total across all catchments, 175 plots and 825 subplots
were sampled in primary forests. Birds: there were two repeat surveys of 15-min
point counts at three sampling points (0, 150 and 300 m) (Extended Data Fig. 1).
Sampling was undertaken between 15 min before dawn and 09:30. Lists of voucher
sound-recordings and images are available for both regions31,32. In total across all
catchments, 1,050 point counts were undertaken in primary forests. Dung beetles:
sampled using pitfall traps (14 cm radius, 9 cm height) baited with 50 g of dung
(80% pig and 20% human) and half filled with a killing solution (5% detergent
and 2% salt). Traps were left for 48 h before inspection. Three traps were placed at
the corners of a 3 m equilateral triangle, repeated at three sampling points (0, 150
and 300 m). In total across all catchments, 1,575 pitfall traps were set in primary
forests (Extended Data Fig. 1).
Defining the biodiversity consequences of forest loss, landscape and within-
forest disturbance. We limit the biodiversity consequences of forest loss to those
that occur in deforested areas themselves, excluding any additional effects on
remaining forests. Landscape disturbance then captures the combined edge, area
and isolation effects that accompany the deforestation process. Within-forest
disturbance refers to anthropogenic disturbance events that are not inevitable
consequences of forest loss or land cover change, including wildfires, hunting
and selective logging. Although often associated with landscape factors, such as
distance from forest edge, within-forest disturbance can occur independently of
changes in forest cover or landscape configuration.
Estimating the conservation value deficit. We used the sum of forest species
presences in primary forest plots to measure the conservation value of a catchment.
In practice, this means that if a forest species occurs on x plots within a catchment,
the species contributes x to the catchment’s conservation value. Total catchment
conservation value is found by summing presences over all forest species. This
measure is equivalent to mean species richness (per unit area) in primary forests
multiplied by primary forest cover. In the absence of disturbance, conservation
© 2016 Macmillan Publishers Limited. All rights reserved
Letter RESEARCH
value should therefore respond linearly to forest cover, with slope equal to mean
species density, de. We term the difference between this linear expectation and a
catchment’s observed conservation value as its conservation value deficit (CVD).
We took a variety of approaches to calculating the CVD, reflecting different
methods of defining forest species, weighting their importance, and calculating de.
Defining forest species. We restricted our analysis to ‘forest species’ to avoid
attributing value to invasive and open-area species. We used three species classi-
fication filters: (i) an automatic filter defined forest species as those that occurred
at least once in a primary forest plot, irrespective of the plot’s disturbance history
(n =​ 1,621 species); (ii) a high basal area (HBA) filter defined forest species to
be those that occurred at least once in plots with a high average basal area (that
is, greater than or equal to the lowest basal area recorded in undisturbed forests
in each region) (n =​ 1,290); and (iii) a convex hull filter where we first applied a
two-dimensional non-metric multidimensional scaling (MDS) to primary and
secondary forest plots based on a stem-size classification (stress =​ 0.14), and then
defined forest species to be those that occurred at least once in plots within the
minimum convex hull of undisturbed primary forest plots on the MDS (n =​  1,140).
Species conservation or functional importance. We used three approaches to
weight species’ importance. First, we assumed that all forest species had a value
equal to 1. Second, we applied a linear weighting to birds according to their range
size and plants according to their wood density. The bird species with the smallest
range size was given a value of 1, and that with the largest range size was given a
value of 0 (vice versa for plants and wood density), with all other species’ values
mapped linearly between these two points. Third, we squared the linear weighting
to give an even higher relative value to species of highest conservation or functional
importance.
There are many important life-history traits that correlate with species’ conser-
vation or functional importance. Our choices were based on a priori knowledge
and the availability of data for diverse tropical taxa. For birds, we chose range size
because it is the single most important predictor of threat status23, especially among
lowland passerines (which make up the majority of our sample) where it is inversely
correlated with other important factors such as population density22. For plants, we
chose wood density because it is the most important size-independent determinant
of carbon storage within individual stems, a strong predictor of carbon stocks
across the biome24,26, and is also linked with other functional properties24 including
drought resistance25. Bird range sizes were extracted from the Birdlife Datazone
(http://www.birdlife.org/datazone/index.html). Wood densities were adapted from
the global wood density database34, using the genus or family average where species
or genus data were unavailable. Lianas were given a nominal value of 0.01.
As part of the broader sensitivity analysis we also undertook the same analy-
sis described above for birds replacing species range size for species mean body
size (body size data was also extracted from Birdlife Datazone). This analysis was
undertaken to determine whether the population density of birds, which is strongly
and inversely correlated with body size, significantly affected results. It did not:
the median estimate of the disturbance impact decreased by just 0.5%, and we do
not report the full results here.
Alternative undisturbed baselines. Estimating de (mean species density in
undisturbed landscapes) requires species distribution data from catchments
with no within-forest or landscape disturbance. As we do not have a set of such
catchments in either region, we took three approaches to calculating de. The first
two approaches rely on the least disturbed catchment in each region. In both
Paragominas and Santarém, this reference catchment had minimal landscape dis-
turbance (>​99% primary forest). However, ground-based observations indicated
that either selective logging or wildfire had affected at least 25% of the sampling
plots within the reference catchments in both regions. We therefore calculated de
as the mean species density over all plots in the reference catchment and, to correct
for within-forest disturbance, as the mean density over only undisturbed reference
catchment plots. Finally, to account for potential biases in underlying (natural)
species distributions, we also calculated de using all undisturbed plots through-
out each region (n =​ 30). This represents a more conservative estimate because it
includes plots in catchments with less than 100% forest cover.
Selecting representative estimates of CVD. Combining the three forest species
selection methods, the three species’ weighting approaches, and the three esti-
mates of de returns 27 estimates of the CVD. For all approaches, we determined
the average CVD with respect to primary forest cover by modelling the catch-
ments’ summed presences with Poisson polynomial generalized linear models.
We selected the best fitting model over all polynomials of degree up to cubics.
To express uncertainty over our estimates of the CVD, in the main text we
present the median relationship between conservation value and forest cover along
with the lower and upper bound range. We excluded from this range the estimate of
de that included disturbed reference catchment plots because it is not reflective of
species density in the absence of disturbance. For the purposes of comparison, we
 RESEARCH Letter
have included these results in Extended Data Fig. 4. The median, lower and upper
bound estimates of the CVD were given by, respectively: the convex hull filter,
linear species weighting, and undisturbed reference catchment plots; the convex
hull filter, no species weighting, and all undisturbed plots; and the high basal area
filter, exponential species weighting, and undisturbed reference catchment plots.
Adjusting for proportionality. Although the number of plots in catchments was
proportional to forest cover, proportionality was not exact because the original
distribution was based on the extent of both primary and secondary forests18.
We therefore corrected sampling effort by calculating the factor required to make
sampling proportional to primary forest cover in each catchment and scaled our
estimates of conservation value accordingly. For each catchment i, this factor is
given by pi/ti , where pi is the proportion of catchment i that is primary forest and
ti is the number of primary forest transects in catchment i.
Extrapolating the CVD. To estimate disturbance impacts throughout Pará, we
divided the state into grid cells approximately equal in size to our study catchments.
We then used Brazil’s 2010 Terraclass product35 to determine the area of each cell
that was deforested, first removing non-forested areas that were covered by water
or tropical savannah. We then calculated each cell’s conservation value by applying
the median, lower and upper bound estimates of the CVD. The disturbance impact
in forest loss equivalent terms for cell i is given by pi −​ (ai −​  ni)vi, where pi, ai, ni
and vi are, respectively, the cell’s primary forest extent, area, non-forest area and
conservation value.
Linking landscape and within-forest disturbance with species distributions
and traits. We investigated the importance of landscape and within-forest distur-
bance at the plot level rather than the catchment level because many disturbance
drivers act at local scales8,13. Variables representing landscape and within-forest
disturbance were based on the analysis of georeferenced 30 m resolution Landsat
TM (Thematic Mapper) and eTM images from 1988 to 2010 in Paragominas and
1990–2010 in Santarém. These were complemented by covariates that represent
natural variation in soil conditions, elevation and slope. A full description of the
data are available elsewhere18. Variable abbreviations match those in Extended
Data Figs 5–7.
Within-forest disturbance. We measured the cumulative extent of canopy distur-
bance36 by calculating the percentage of the remaining primary forest in a 1 km
buffer around each plot that had never been classified as disturbed (undisturbed
primary forest, UPF). We also included two measures of the frequency of distur-
bance within plots: the number of times the plot was logged (NL) and the number
of times the plot was burnt (NB) in visual inspections of satellite images or field
observations.
Landscape disturbance. We used two landscape configuration measures: the density
of forest-agriculture edges (ED) and the percentage of primary and secondary
(>​10 years old) forest cover (FC) in 1 km buffers around plots. We used two meas-
ures of landscape history37: the deforestation curvature profile (DC) and the land-
use intensity profile (LI) in 500 m buffers around plots.
Natural environmental covariates. We used soil samples and digital elevation mod-
els to derive covariates reflecting natural conditions. Soil variables were based on
average values from five 30 cm deep soil profiles in each plot, and include acidity
(pH), clay content (Cl), and carbon stock (Ca). We applied a 100 m buffer around
each plot in a digital elevation model to calculate mean plot elevation (El) and
slope (Sl).
Linking landscape and within-forest disturbance with species distributions and
traits. We used Random Forests (RF), a decision-tree classification methodology,
to identify species that are well-modelled by our data and to rank the importance
of individual variables in accounting for species distributions. RF was adapted for
spatial autocorrelation within catchments using a modified ‘residual autocorrela-
tion’ approach38. The fit of the RF models and their predictive performance was
measured using area under receiver–operator curves (AUC)39. AUC evaluates the
ability of models to correctly predict higher probability of occurrence where spe-
cies are present than where they are absent. An AUC value of 1 indicates perfect
discrimination; a value of 0.5 suggests predictions no better than random. We
performed multiple cross-validations to evaluate model predictive performance.
For each species, data from each study catchment were used in turn as test data for
models built with data from the other catchments. The cross-validated AUC value,
AUCcv, was calculated as the average AUC value over all cross-validation tests for
each species. Species present on a minimum of three transects and with a summed
© 2016 Macmillan Publishers Limited. All rights reserved
AUCcv ≥​ 0.6 over all variables were classified as well-modelled and included in the
analyses (31% of species). The importance of a variable was measured as its mean
AUCcv over all well-modelled species.
Models included the within-forest disturbance, landscape disturbance and nat-
ural environment covariates described above. Given multicollinearity, we selected
two variables from each group using three variable-selection methods: (i) we
selected variables that we hypothesized to have the greatest influence on species’
presences (hypothesis-driven selection); (ii) we used principal component analysis
(PCA) on the full set of variables in each group and selected the highest loaded
variable on the first two principal axes (PCA selection); and (iii) we ran RF on the
full set of variables and selected the two highest ranked in each group (step-wise
selection). Results for each method are shown in Extended Data Figs 5–7.
Next, we used RF to determine species’ partial responses along disturbance
gradients (Figs 3 and 4 and Extended Data Fig. 8). These partial responses give the
relative odds (exp(logit(p) −​  mean(logit(p)), where p is the probability of species’
presence and logit is ln(p/(1 −​  p))) of detecting each species along a single variable
gradient, holding all other variables constant. For this analysis we selected the
landscape and within-forest disturbance variables that were most frequently ranked
highest in their group across the three variable selection methods.
We then used latent trajectory analysis (LTA), which groups species’ partial
responses into homogenous classes, to characterize the main types of response to
the selected variables. We built models with up to eight classes and selected that
with the lowest Bayesian information criterion (BIC) score. LTAs were carried out
in R package ‘lcmm’ http://cran.r-project.org/web/packages/lcmm/lcmm.pdf. In
Figs 3 and 4, we show the LOWESS smoothed response of each species class along
the associated disturbance gradient, with bandwidth set to 0.75.
Finally, we investigated the relationship between the disturbance sensitivity of
species classes, as determined by LTA, and species traits. To undertake this analysis,
we defined a metric that represents the propensity of species classes to be detected
along the variable gradients, which thus provides a measure of the sensitivity of
the class to disturbance. The measure is:
u m
hc (x) = ∫ (x − m)dc(x)dx − ∫ (m − x)dc(x)dx
m l
where m, l and u are, respectively, the gradient’s mid-point and lower and upper
bounds, and dc(x) is the relative odds of detecting species class c at point x on the
gradient, as determined by RF. We scaled hc to lie between +​1 and −​1. Values of
hc close to 1 indicate that species class c is much more likely to be detected at the
maximum than minimum extreme of the gradient, values close to −​1 indicate that
species class c is much more likely to be found at the minimum than maximum
extreme. Values near 0 indicate that species class c is equally likely to be detected at
either extreme. We tested the relationship between hc and species’ traits by fitting
polynomial models weighted by group size. In all cases, the response variable was
the average value of the species trait over all species in each class. We investigated
polynomial fits up cubics and selected that with the lowest BIC score.
31. Lees, A. C. et al. One hundred and thirty-five years of avifaunal surveys around
Santarem, central Brazilian Amazon. Rev. Bras. Ornitol. 21, 16–57 (2013).
32. Lees, A. C. et al. Paragominas: a quantitative baseline inventory of an eastern
Amazonian avifauna. Rev. Bras. Ornitol. 20, 93–118 (2012).
33. Barrio, J. Hunting pressure on cracids (Cracidae: Aves) in forest concessions in
Peru. Rev. Peru. Biol. 18, 225–230 (2011).
34. Zanne A. E. et al. Data from: Towards a worldwide wood economics spectrum.
Dryad Digital Repository. http://dx.doi.org/10.5061/dryad.234 (2009).
35. Instituto Nacional de Pesquisas Espaciais (INPE). Terraclass data 2010;
available at http://www.Inpe.Br/cra/projetos_pesquisas/terraclass2010 (2013).
36. Souza, C. M. Jr. et al. Ten-year landsat classification of deforestation and forest
degradation in the Brazilian Amazon. Remote Sens. 5, 5493–5513 (2013).
37. Ferraz, S. F. D., Vettorazzi, C. A. & Theobald, D. M. Using indicators of
deforestation and land-use dynamics to support conservation strategies:
A case study of central Rondonia, Brazil. For. Ecol. Manage. 257, 1586–1595
(2009).
38. Crase, B., Liedloff, A. C. & Wintle, B. A. A new method for dealing with residual
spatial autocorrelation in species distribution models. Ecography 35, 879–888
(2012).
39. Pearce, J. & Ferrier, S. Evaluating the predictive performance of habitat models
developed using logistic regression. Ecol. Model. 133, 225–245 (2000).
 Letter RESEARCH

Extended Data Figure 1 | Study design. a, The location of Paragominas and Santarém within Pará. b, c, The distribution of study catchments (n = 36)
within Paragominas and Santarém, respectively. d, The distribution of study plots (n = 175) in example catchments spanning the gradient of primary
forest. Selected catchments are shown in red in a and b. e, Sampling design within each plot.

© 2016 Macmillan Publishers Limited. All rights reserved

 RESEARCH Letter

Extended Data Figure 2 | Richness of forest species. a–c, The richness Paragominas (green) and Santarém (orange). Panels show the convex
of forest species in secondary forests (SF), pastures (PA), and mechanised hull (a), automatic (b) and high basal area filters (c) used to classify forest
agricultural lands (AG) relative to the average richness of forest species species (see Methods).
in all undisturbed and disturbed primary forests (dashed line) in

© 2016 Macmillan Publishers Limited. All rights reserved

 Letter RESEARCH

Extended Data Figure 3 | Conservation value of primary forests
measured by individual taxa. a–d, Estimates of conservation value in
the Paragominas (circles) and Santarém (triangles) study regions from
large-stemmed plants (a) small-stemmed plants (b) birds (c) and
dung beetles (d). Dashed lines show expectations without disturbance.
Grey lines show all regressions, with the black solid line showing the
median response (see Methods). Values were standardized across study
regions and taxa. There was no significant difference between taxa in the
median estimate (F3;117 =​  1.36, P =​  0.26, ANCOVA).

© 2016 Macmillan Publishers Limited. All rights reserved

 RESEARCH Letter

Extended Data Figure 4 | Range of conservation value estimates using three alternative sets of reference plots. Mean species density (de) is measured
by: all disturbed and undisturbed plots in the least disturbed reference catchments (grey shaded region), all undisturbed plots throughout a region (green
shaded region), and undisturbed plots in the reference catchments (purple shaded region). See Methods for details.

© 2016 Macmillan Publishers Limited. All rights reserved

 Letter RESEARCH

Extended Data Figure 5 | The importance of hypothesis selected variables.
a–h, Species AUCcv values for each variable in Paragominas (a, c, e, g) and
Santarém (b, d, f, h) for large-stemmed plants (a, b), small-stemmed plants
(c, d), birds (e, f) and beetles (g, h). Variable importance was measured by
the mean AUCcv over all well-modelled species (see Methods). Variable
colours denote group membership: green, orange and blue represent
landscape disturbance, within-forest disturbance and natural variables,
respectively (see Methods for variable descriptions). Letters show the
results for multiple pair-wise comparisons of group means using Tukey’s
range test. Variables which do not share a letter have significantly different
mean importance (P <​  0.05).

© 2016 Macmillan Publishers Limited. All rights reserved

 RESEARCH Letter

Extended Data Figure 6 | The importance of PCA selected variables.
a–h, Species’ AUCcv values for each variable in Paragominas (a, c, e, g)
and Santarém (b, d, f, h) for large-stemmed plants (a, b), small-stemmed
plants (c, d), birds (e, f) and beetles (g, h). Variable importance was
measured by the mean AUCcv over all well-modelled species (see Methods).
Variable colours denote group membership: green, orange and blue
represent landscape disturbance, within-forest disturbance and natural
variables, respectively (see Methods for variable descriptions). Letters
show the results for multiple pair-wise comparisons of group means using
Tukey’s range test. Variables which do not share a letter have significantly
different mean importance (P <​  0.05).

© 2016 Macmillan Publishers Limited. All rights reserved

 Letter RESEARCH

Extended Data Figure 7 | The importance of step-wise selected variables.
a–h, Species’ AUCcv values for each variable in Paragominas (a, c, e, g)
and Santarém (b, d, f, h) for large-stemmed plants (a, b), small-stemmed
plants (c, d), birds (e, f) and beetles (g, h). Variable importance is
measured by the mean AUCcv over all well-modelled species (see
Methods). Variable colours denote group membership: green, orange
and blue represent landscape disturbance, within-forest disturbance and
natural variables, respectively (see Methods for variable descriptions).
Letters show the results for multiple pair-wise comparisons of group
means using Tukey’s range test. Variables which do not share a letter have
significantly different mean importance (P <​  0.05).

© 2016 Macmillan Publishers Limited. All rights reserved

 RESEARCH Letter

Extended Data Figure 8 | Responses of small-stemmed plants and in Paragominas (a, c, e, g) and Santarém (b, d, f, h) (see Methods). Species
dung beetles to disturbance. a–h, The odds of detecting small-stemmed groups, shown by different coloured lines, are composed of species with
plants (a–d) and dung beetles (e–h) species groups along gradients of similar disturbance responses (see Methods). Line thickness represents the
landscape disturbance (a, b, e, f) and within-forest disturbance (c, d, g, h) relative size of the groups.

© 2016 Macmillan Publishers Limited. All rights reserved

 Letter RESEARCH

Extetnded Data Table 1 | Policy interventions used to reduce deforestation and their effect on disturbance

© 2016 Macmillan Publishers Limited. All rights reserved

 RESEARCH Letter

Extended Data Table 2 | Forest loss and disturbance in Pará and its areas of endemism

a–c, The loss of primary forest conservation value from forest loss and forest disturbance in forest loss-equivalent terms across ~​50  km2 cells covering all land in Pará (a), private lands and
sustainable use reserves only (b), and private lands only (c). Disturbance losses are calculated using the median estimate of conservation value with the lower and upper bound range in parentheses
(see Methods). Area is the total area of the region in km2. Forest area gives the area of the region that was or is primary forest cover in km2. Forest loss gives the total loss of primary forest in km2.
Disturbance gives the loss of conservation value due to disturbance in km2. Relative gives the disturbance-mediated loss of conservation value relative to that from forest loss.

© 2016 Macmillan Publishers Limited. All rights reserved